Proposal of a New Approach Using Deep Learning for QR Code

Embedding

Kanaru Kumabuchi

and Hiroyuki Kobayashi

Osaka Institute of Technology University, Osaka, Japan

Keywords:

Deep Learning, Image Hiding, Image Processing.

Abstract:

The purpose of this research is to enhance the technique of embedding QR codes into arbitrary images using

deep learning. Previous approaches faced the issue of compromising the quality when embedding QR codes

into arbitrary images. We address this problem by proposing a deep learning model and learning method

that can improve the quality of embedded images and accurately recover QR codes. Speciﬁcally, we design

a new model using deep learning that embeds QR codes into images while minimizing the degradation of

image quality. The effectiveness of the proposed model and learning method is validated through experiments,

demonstrating the enhancement of image quality in the embedded images and accurate QR code recovery.

1 INTRODUCTION

In recent years, with the widespread use of the inter-

net, exchanging information and communication has

become convenient. However, on the other hand, the

leakage of personal information and organizational

assets has become a signiﬁcant problem. As a coun-

termeasure, there is a technique called steganography.

Steganography is the art of concealing one piece of

digital data (audio, images) within another piece of

digital data.

In a previous research(Kumabuchi and Kobayashi,

2022), two models were created using deep learning:

one to embed QR codes into images and the other to

restore QR codes from images with embedded QR

codes. However, there was a signiﬁcant issue with

embedding QR codes into images, as it substantially

compromised the quality of the original images. In

this research, similar to the previous study, we aim

to create new Encoder and Decoder models to embed

QR codes into images and restore them without com-

promising the quality of the original images. We pro-

pose and evaluate a model capable of achieving this

goal

In a previous research, we referred to the model

proposed by Simon J

egou(J

egou et al., 2017). for

semantic segmentation, which improved upon the

Unet(Ronneberger et al., 2015) model, and used it as

https://orcid.org/0009-0004-0181-4185

https://orcid.org/0000-0002-4110-3570

a basis for our work. In this study, we further re-

ﬁned that model to devise a method for embedding

QR codes while preserving their distinctive features.

2 PRINCIPLE

In this PRINCIPLE, the embedding and restoration

procedures of the QR code are explained with the aid

of Figure1, along with the learning steps.

1. Input a three-channel image and a one-channel

QR code into two separate models.

2. Concatenate the two output feature maps at an in-

termediate layer and input them into the Conca-

tImageModel.

3. In the QR code embedding model, train with the

three-channel image as the ground truth.

4. Pseudo-image and normalize it before inputting it

into the Restoration Model.

5. Train the Restoration Model using the output im-

age as input and the one-channel QR code image

as the ground truth.

6. Next, compute the loss for both models using the

Mean Squared Error (MSE) from the following

equation (1), and then calculate the weighted loss

using the following equation (2).

7. Use the computed loss values to update the

weights of both models.

342

Kumabuchi, K. and Kobayashi, H.

Proposal of a New Approach Using Deep Learning for QR Code Embedding.

DOI: 10.5220/0012238900003543

In Proceedings of the 20th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2023) - Volume 1, pages 342-345

ISBN: 978-989-758-670-5; ISSN: 2184-2809

Figure 1: Overall Structure.

Loss = mse =

· M

∑

n=1

∑

n=1

{y(i, j) − x(i, j)}

(1)

loss = α · Loss

Encoder

+ β ·Loss

Decoder

(2)

3 MODEL CREATION

3.1 Datasets

The dataset used in this research consists of indoor

images shown in Figure 2 and images generated us-

ing Python’s QR code library as depicted in Figure 3.

For the training process, 8000 images and QR codes

were utilized for each category, and an additional set

of 1000 images was reserved for testing purposes.

Figure 2: Indoor image.

3.2 Conventional Model

In this research, we aimed to enhance the perfor-

mance by utilizing an improved model compared to

the conventional approach. Before explaining the

model used in this study, let’s ﬁrst describe the con-

ventional model. The conventional model combines

Figure 3: QR Image.

a 3-channel image with a QR code and inputs them

together into a single Encoder-Decoder model. How-

ever, in our model, we take a different approach by

performing dimensionality reduction separately for

the QR code and the 3-channel image using distinct

models. This enables us to embed the QR code into

the 3-channel image while preserving its distinctive

features. +

Input Image Input Image

Figure 4: Conventional Model.

3.3 Model Structure

The overall structure of this research is as shown

in Figure 1, comprising two models. The Encoder

Model in the ﬁgure serves as a model for embedding

QR codes. On the other hand, the Decoder Model is

used for recovering QR codes embedded in images.

In the following subsections, I will provide a detailed

explanation of the structure of each model.

3.3.1 Encoder Model

The structure of the Encoder model is an Encoder-

Decoder architecture. In this architecture, both three-

channel images and QR codes are input into the same

Encoder shown in Figure 5c. Dimensional compres-

Proposal of a New Approach Using Deep Learning for QR Code Embedding

343

Conv2D

Batch Norm

Relu

Dropout

Input

Output

(a) Dence parts.

Dence

Parts

Dence

Parts

Dence

Parts

Dence

Parts

Input

Output

(b) Dence block.

Input

Image

Input

Dence

Parts

Dence

Parts

Dence

Parts

Dence

Parts

Concat

Image

Dence

Parts

Dence

Parts

Embedded

Image

Dence

Parts

Dence

Parts

Dence

Parts

Dence

Parts

Embedded

Image

Restoration

Image

(d) Decoder model.

Figure 5: Detail Structure.

sion is performed, and the intermediate layers are con-

catenated before decoding is done. The model struc-

ture consists of Dense Blocks, as depicted in Figure

5b. Inside the Dense Block, there are convolutional

layers, batch normalization layers, ReLU layers, and

Dropout layers as shown in Figure 5a. The use of skip

connections for all layers prevents the vanishing gra-

dient problem

3.3.2 Decoder Model

The Decoder model aims to restore a QR code from

an image containing an embedded QR code. The De-

coder model has a simple structure, consisting of four

connected Dense blocks as shown in the Figure 5d.

4 EXPERIMENTAL

In this EXPERIMENTAL, we conducted a 200-epoch

training using the learning procedure described in the

principles and the model illustrated in Figure1. Sub-

sequently, we utilized the trained model to compare

the output results with those obtained from the con-

ventional model, thus examining the differences be-

tween them

4.1 Results of Conventional Models

After training the model using the architecture shown

in Figure 5, we obtained the results for the test im-

ages, as shown in Figure 7. However, it is evident

that while the conventional model can restore the QR

code, the images with embedded QR codes result in a

loss of image quality in the input images

(a) Input Image. (b) Input QRcode.

Figure 6: Conventional model result.

4.2 Results of this Research Model

After training the model using the procedure shown in

Figure 1, we obtained results for test images as shown

in Figures 8 and 9. In comparison to the results of the

conventional model depicted in Figure 6, it was con-

ﬁrmed that not only can QR codes be restored, but

they can also be embedded more clearly into the im-

ages. Furthermore, the presence of areas in Figure

8 where embedding is not complete is believed to be

due to high brightness values.

5 CONCLUSION

In this research, we developed new Encoder and De-

coder models to improve the performance of both im-

age embedding and QR code restoration. As a result,

we were able to obtain output images with embed-

ded QR codes that closely resembled the input im-

ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics

344

(a) Input Image. (b) Input QRcode.

Figure 7: This research model result.

(a) Input Image. (b) Input QRcode.

Figure 8: This research model result.

ages, and the restored QR codes were in a readable

state.

REFERENCES

egou, S., Drozdzal, M., Vazquez, D., Romero, A., and Ben-

gio, Y. (2017). The one hundred layers tiramisu: Fully

convolutional densenets for semantic segmentation. In

Proceedings of the IEEE conference on computer vi-

sion and pattern recognition workshops, pages 11–19.

Kumabuchi, K. and Kobayashi, H. (2022). Improving the

performance of qr code embedding in arbitrary images

using deep learning. volume 2022, pages 1P1–Q08.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-

net: Convolutional networks for biomedical image

segmentation. In Medical Image Computing and

Computer-Assisted Intervention–MICCAI 2015: 18th

International Conference, Munich, Germany, October

5-9, 2015, Proceedings, Part III 18, pages 234–241.

Springer.

Proposal of a New Approach Using Deep Learning for QR Code Embedding

345